“Your Word is my Command”: Google Search by Voice: A Case Study

Abstract

An important goal at Google is to make spoken access ubiquitously available. Achieving ubiquity requires two things: availability (i.e., built into every possible interaction where speech input or output can make sense) and performance (i.e., works so well that the modality adds no friction to the interaction).This chapter is a case study of the development of Google Search by Voice -a step toward our long-term vision of ubiquitous access. While the integration of speech input into Google search is a significant step toward more ubiquitous access, it has posed many problems in terms of the performance of core speech technologies and the design of effective user interfaces. Work is ongoing and no doubt the problems are far from solved. Nonetheless, we have at the minimum achieved a level of performance showing that usage of voice search is growing rapidly, and that many users do indeed become repeat users.

Asynchronous stochastic optimization for sequence training of deep neural networks

“…An empirical study of ASGD for sequence training was conducted for a small and a large Voice Search task [21]. The small size of the Icelandic dataset allows for a thorough evaluation of the different issues and hyper-parameters while the large scale English dataset demonstrates the scalability of the algorithm.…”

Multiframe deep neural networks for acoustic modeling

“…The goal of this paper is to look at the complexity/accuracy tradeoff of such a system. To explore this tradeoff, we trained a collection of systems of various complexities on two datasets: US English Voice Search [5] and voice typing, and Iberian Portuguese Voice Search, by varying the width of the hidden layers of the DNN acoustic model. The rest of the system was kept fixed: the frontend consists of 40 log-filterbank energies computed every 10 ms, stacked 20 frames in the past and 5 frames in the future to limit latency.…”

Mobile music modeling, analysis and recognition

“…As a baseline, we have used a standard speech recognition front end computing perceptual linear predictive (PLP) features [8]. Thirteen cepstral coefficients are computed every 10 ms over 25 ms windows, and the energy coefficient C0 is discarded, which decreases the sensitivity to changes in volume.…”